MR-based Radiomics And Machine Learning For Prediction Of Pathological Grade,Microvascular Invasion And Ki-67 In Hepatocellular Carcinoma

Part I.MR-based radiomics and machine learning for prediction of pathological grade in hepatocellular carcinomaBackground&Objective:The pathological grade of hepatocellular carcinoma（HCC）is associated with the prognosis.To develop and validate the machine learning models based on MR radiomics features for preoperative prediction of pathological grade of HCC.Methods:Two hundred eighty-one patients with a diagnosis of HCC and indications for hepatic resection were included,and assigned to train（n=210）and test（n=71）set.The clinical and MR data were retrospectively analyzed,and the following features were evaluated:maximum diameter,tumor margins,tumor capsule,peritumoral enhancement on arterial phase（AP）,peritumoral hypointensity on hepatobiliary phase（HBP）,intratumoral fat,intratumoral necrosis and intratumoral haemorrhage.Radiomics features were extracted from Gd-EOB-DTPA enhanced MR T₁WI precontrast phase,arterial phase（AP）,portal vein phase（VP）,delayed phase（DP）and hepatobiliary phase（HBP）images in regions of the intratumoral areas.The maximum relevance minimum redundancy（m RMR）,least absolute shrinkage and selection operator（LASSO）algorithm,and random forest（RF）-based recursive feature elimination（RFE）were applied to select radiomics features.Stepwise logistic regression based on Akaike information criterion（AIC）was used to select clinical-radiological features.The Support Vector Machines（SVM）,Multi-layer Perceptron（MLP）,Extra Trees Classifier（ETC）and e Xtreme Gradient Boosting Classifier（XGBC）algorithms were applied to develop models.Discrimination and calibration of the models were assessed.Results:The area under receiver-operating-characteristic curve（AUC）of clinical-radiological models were 0.608（SVM）,0.601（MLP）,0.597（ETC）and 0.615（XGBC）in the test set,respectively.The AUC of radiomics models were 0.728（SVM）,0.732（MLP）,0.725（ETC）and 0.759（XGBC）in the test set,respectively.The AUC of the combined models based on MR radiomics features and clinical-radiological features were 0.762（SVM）,0.701（MLP）,0.711（ETC）and 0.719（XGBC）in the test set,respectively.Conclusion:The machine learning models（SVM,ETC,MLP and XGBC）based on the multi-phase MR radiomics features showed favourable predictive efficiency for preoperative prediction of pathological grade of HCC.The ETC and XGBC models based on MR radiomics features demonstrated better performance than other ones.The clinical-radiological features showed poor performance for predicting pathological grade of HCC.Part II.MR-based radiomics and machine learning for prediction of microvascular invasion in hepatocellular carcinomaBackground a& Objective: Microvascular invasion（MVI）is a high-risk factor for early postoperative recurrence and poor prognosis of patients with HCC.Preoperative prediction of MVI is important for surgery strategy making.To develop and validate the machine learning models based on MR radiomics features for preoperative prediction of MVI in HCC.Methods: Two hundred thirty-two patients with a diagnosis of HCC and indications for hepatic resection were included,and assigned to train（n=174）and test（n=58）set.The clinical and MR data were retrospectively analyzed,and the following features were evaluated: maximum diameter,tumor margins,tumor capsule,peritumoral enhancement on AP,peritumoral hypointensity on HBP,intratumoral fat,intratumoral necrosis and intratumoral haemorrhage.Radiomics features were extracted from Gd-EOB-DTPA enhanced MR T1 WI precontrast phase,arterial phase,portal vein phase,delayed phase and hepatobiliary phase images in regions of the intratumoral and peritumoral areas.The m RMR and LASSO algorithms were applied to select MR radiomics features related to MVI.Stepwise logistic regression based on AIC was used to select clinical-radiological features.The SVM,LR,MLP and ETC algorithms were used to develop models.Discrimination and calibration of the models were assessed.Results: The AUC of clinical-radiological models were 0.746（SVM）,0.762（LR）,0.750（MLP）and 0.790（ETC）in the test set,respectively.The AUC of radiomics models were 0.751（SVM）,0.726（LR）,0.736（MLP）and 0.778（ETC）in the test set,respectively.The AUC of the combined models based on MR radiomics features and clinical-radiological features were 0.766（SVM）,0.816（LR）,0.816（MLP）and 0.826（ETC）in the test set,respectively.Radscore was significantly different between MVI-present and MVI-absent patients（p<0.05）.The AUC of the combined models based on Radscore and clinical-radiological features were 0.767（SVM）,0.781（LR）,0.789（MLP）and 0.785（ETC）in the test set,respectively.Conclusion: The machine learning models（SVM,LR,ETC and MLP）based on the intratumoral and peritumoral multi-phase MR radiomics features showed favourable predictive efficiency for preoperative predicting MVI.The higher the Radscore,the higher the risk of MVI.The ETC model based on MR radiomics features and clinical-radiological features demonstrated better performance than other ones.Part III.MR-based radiomics and machine learning for prediction of Ki-67 in hepatocellular carcinomaBackground & Objective: Ki-67 is a biomarker of tumor proliferation.The higher Ki-67 index indicates the higher invasive and poor prognosis of HCC.To develop and validate the machine learning models based on MR radiomics features for preoperative prediction of Ki-67 expression in HCC.Methods: Two hundred and ten patients with a diagnosis of HCC and indications for hepatic resection were included,and assigned to train（n=157）and test（n=53）set.The clinical and MR data were retrospectively analyzed,and the following features were evaluated: maximum diameter,tumor margins,tumor capsule,peritumoral enhancement on AP,peritumoral hypointensity on HBP,intratumoral fat,intratumoral necrosis and intratumoral haemorrhage.Radiomics features were extracted from Gd-EOB-DTPA enhanced MR T1 WI precontrast phase,arterial phase,portal vein phase,delayed phase and hepatobiliary phase images in regions of the intratumoral areas.The m RMR and LASSO algorithms were used to select radiomics features related to Ki-67 expression in HCC.Stepwise logistic regression based on AIC was used to select clinical-radiological features.The LR,SVM,MLP and ETC algorithms were applied to develop models.Discrimination and calibration of the models were assessed.Results: The AUC of the clinical-radiological models were 0.790（LR）,0.755（SVM）,0.781（MLP）and 0.785（ETC）in the test set,respectively.The AUC of the radiomics models were 0.810（LR）,0.817（SVM）,0.814（MLP）and 0.808（ETC）in the test set,respectively.The AUC of the combined models based on radiomics features and clinical-radiological features were 0.839（LR）,0.827（SVM）,0.864（MLP）and 0.813（ETC）in the test set,respectively.Radscore was significantly different between low Ki-67 and high Ki-67 patients（p<0.05）.The AUC of the combined models based on Radscore and clinical-radiological features were 0.813（LR）,0.825（SVM）,0.803（MLP）and 0.824（ETC）in the test set,respectively.Conclusion: The machine learning models（LR,SVM,ETC and MLP）based on multiphase MR radiomics features can effectively predict the expression of Ki-67 in HCC.The higher the Radscore,the higher expression of Ki-67.The MLP model based on MR radiomics features and clinical-radiological features demonstrated better performance than other ones.